ChangeMambaVision: Adapting MambaVision for Building Change Detection

Change Detection (CD) is important for analysis of land cover change within a geographical area. Accurate manual classification and labeling of change areas is difficult, and change results are used for subsequent data analysis. Therefore, there is a need for highly accurate automated methods for CD. The recent Mamba architecture has seen growing usage in vision applications due to its lower time complexity compared to Transformer, which provides better scalability for very high-resolution (VHR) imagery. MambaVision is a novel architecture that incorporates Mamba blocks alongside convolution blocks to enhance feature extraction of finer details, and Transformer blocks to enhance the capture of global dependencies. Mam-baVision has shown reported state-of-the-art performance on general vision tasks, as well as achieving unparalleled throughput compared to other vision backbones. In this paper, we propose ChangeMambaVision, where we adapt the MambaVision backbone with minimal modifications for Change Detection with a lightweight ChangeFormer decoder to complement its strengths. We show that ChangeMambaVision on the Base capacity achieves better performance than the current state-of-the-art (CDMamba) on two Building CD datasets, namely the LEVIR-CD (+0.56 F1 / +0.92 IoU) and WHU-CD (+0.49 F1 / +0.77 IoU), while having faster throughput (68.5% increase vs. CDMamba) and lower training memory costs (76.68% of CDMamba training memory). Code, training logs and trained weights are available in https://github.com/regencode/ChangeMambaVision.

Authors:

Thomas Gozalie, Andrea Stevens Karnyoto, Bens Pardamean, Edy Irwansyah